Research On Point Cloud Classification Based On Graph Convolution

With the wide application of 3D sensors such as LiDAR and RGBD camera in robots and driverless technology,the research based on 3D point cloud data has also achieved some development,such as object classification,partial segmentation and scene semantics and so on.The object classification of point clouds is the most basic research in the research of point cloud data,so it has become a hot and difficult direction in the research of point cloud data.The current research on disordered point cloud classification is mainy converts point cloud data into a regular 3D voxel grid or image sets(projecting points from different angles in 2D).This involves extensive data addition,preprocessing and heavy computational problems.In addition,most of these point cloud classification methods extract features in an isolated manner,ignoring the location of neighboring points and their geometric features.This will affect the classification effect and the expand research of 3D point clouds.In order to solve the problems of limited methods and point features have been abstracted in an independent and isolated manner,this paper studies a point cloud classification method based on graph convolution,and constructs and implements local point cloud classification based on graph convolution and global point cloud classification based on graph convolution.This paper uses gra ph convolution,combining new graph pooling strategies for more efficient and accurate object classification of 3D point cloud.The specific work of the paper is as follows:In order to solve the problem of point features have been abstracted in an independent and isolated manner,ignoring the relative location of neighboring points as well as their features,this paper proposes a local point cloud classification model based on graph convolution.The model encodes point cloud data.The graph convolution is performed on the nearest neighbor graph of the point neighborhood constructed by the graph code.To avoid the high computational cost caused by graph coarsening or clustering,feature merging is performed by a multi-resolution pooling strategy.The model not only changes the current point cloud data to learn features in an independent and isolated manner,but also as a lightweight point cloud classification model avoids the problem of heavy data preprocessing in the current point cloud classification methods.In order to solve the over-fitting problem of local point cloud classification model based on graph convolution,a global point cloud classification model based on graph convolution is proposed.The model solves the problem of overfitting by seeking a representation of the order of the points and the rotation invariant.In this paper,the new graph coding method is combined with the global pooling method,and the new equilibrium loss function and more complex classification framework are utilized,which greatly avoids over-fitting.Although the model structure is more complicated,it has the advantages of fast convergence,high accuracy,and good robustness in classification effect.On the standard dataset ModelNet40,the proposed local point cloud classi fication model based on graph convolution and the global point cloud classification model based on graph convolution are experimentally verified.Experiments show that both methods proposed in this paper show a good classification effect.